A noise reduction system for a voice mouse

Abstract

The application discloses a noise reduction system for a voice mouse, relates to the technical field of audio noise reduction, and solves the technical problem of low recognition accuracy of the voice mouse; a voice acquisition module is used for acquiring voice signals; and the voice signals are sent to a voice noise reduction module; the voice noise reduction module performs noise reduction processing on the first voice signal and the second voice signal respectively, and obtains a first clean voice signal and a second clean voice signal; the first clean voice signal and the second clean voice signal are synthesized to obtain a clean voice signal; the voice signals are subjected to noise reduction processing, the data traffic in data communication can be reduced, and the integrity of the data is ensured; the clean voice signal is compensated according to an AGC algorithm, reverse gain is used when the user's voice exceeds the acquisition limit, audio distortion is prevented, and the fidelity and accuracy of the audio are improved.

Description

Technical Field

[0001] This invention belongs to the field of smart devices and relates to noise reduction technology, specifically a noise reduction system for a voice mouse. Background Technology

[0002] Voice-activated mice, with their voice recognition capabilities, can not only transcribe user speech into text for input but also recognize user voice commands to control the computer. They also allow for intelligent button control, significantly improving the convenience and efficiency of computer operation. As the user base and application scenarios of voice-activated mice expand, they are attracting increasing attention.

[0003] Currently, there are two types of voice-enabled mice on the market. The first type is based on a single microphone and lacks voice noise reduction. In noisy or chaotic environments, the accuracy of audio recognition is significantly reduced. The second type is based on dual microphones and uses a host computer for noise reduction (utility model patent CN210428381U). While it has noise reduction, it lacks AGC (Automatic Gain Control). Due to bandwidth limitations in the communication protocol, audio transmission needs to be compressed before being sent to the host computer for model algorithm noise reduction. This method cannot handle audio distortion at the mouse end, and the loss of audio information caused by compression makes it difficult to accurately capture audio in noisy or loud environments, thus affecting recognition accuracy. It also significantly reduces the accuracy of recognizing homophones and the recording experience, impacting the user experience to some extent.

[0004] To address this, a noise reduction system for a voice mouse is proposed. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a noise reduction system for a voice mouse. This noise reduction system for a voice mouse solves the problems of existing technologies, which, although possessing noise reduction capabilities, lack AGC (Automatic Gain Control). Furthermore, due to bandwidth limitations in communication protocols, audio transmission requires compression before being sent to the host computer for model algorithm noise reduction. These systems cannot handle audio distortion at the mouse end, suffer from audio information loss due to compression, and fail to accurately capture audio in noisy and loud environments, thus affecting recognition accuracy. The system also significantly reduces the accuracy of homophone recognition and the recording experience, impacting the user experience to some extent.

[0006] To achieve the above objectives, an embodiment of the first aspect of the present invention provides a noise reduction system for a voice mouse, comprising a voice acquisition module, a voice noise reduction module, and a voice recognition module;

[0007] The voice acquisition module is used to acquire voice signals; wherein, the voice signals include a first voice signal and a second voice signal;

[0008] And the voice signal is sent to the voice noise reduction module;

[0009] The speech noise reduction module is used to process the speech signal to obtain a clean speech signal, including the following steps:

[0010] Step 1: The speech noise reduction module performs noise reduction processing on the first speech signal and the second speech signal respectively to obtain a first clean speech signal x1(n) and a second clean speech signal x2(n);

[0011] Step 2: Combine the first clean speech signal and the second clean speech signal to obtain the clean speech signal x(n);

[0012] Step 3: Compensate the clean speech signal according to the AGC algorithm;

[0013] The speech recognition module is used to convert clean speech signals into text data.

[0014] Preferably, the voice acquisition module includes a first voice acquisition unit and a second voice acquisition unit;

[0015] The first voice acquisition unit is used to acquire the first voice signal;

[0016] The second voice acquisition unit is used to acquire the second voice signal.

[0017] Preferably, the speech noise reduction module performs noise reduction processing on the first speech signal, including the following steps:

[0018] S1: Extract the first M frames of the first speech signal as the first initial speech signal; where M is an integer greater than 0 and less than 20;

[0019] S2: Input the first initial speech signal and the first speech signal data into the noise estimation model to obtain the first initial gain G1(i,k);

[0020] S3: Divide the first speech signal into P segments and count the energy of each segment;

[0021] The speech noise reduction module sets a quantile value and marks speech signals with energy lower than the quantile value as noise signals.

[0022] S4: The speech noise reduction module inputs the speech signal marked as noise into the noise analysis model to obtain the noise probability; and marks the speech signal with the highest noise probability as the second initial speech signal;

[0023] S5: Input the second initial speech signal and the first speech signal data into the noise estimation model to obtain the second initial gain G2(i,k);

[0024] S6: Obtain the clean speech x1(n) after denoising the first speech signal based on the first initial gain and the second initial gain.

[0025] Preferably, the noise estimation model is used to obtain the gain G(i, k);

[0026] The principle of the noise estimation model includes:

[0027] y(n) = x(n) + d(n)

[0028] Where x(n) represents the clean speech signal, d(n) represents the noise signal, and y(n) represents the signal collected by the microphone;

[0029] Performing a short-time Fourier transform on y(n) = x(n) + d(n), we obtain:

[0030] Y(i,k)=X(i,k)+D(i,k)

[0031] Where Y(i,k), X(i,k) and D(i,k) represent the frequency domain signals of y(n), x(n) and d(n) respectively, i represents the i-th time domain frame, and k represents the frequency point;

[0032] G(i,k)=|X(i,k)| / |Y(i,k)|.

[0033] Preferably, in this embodiment, an artificial intelligence model is trained based on noise classification features to obtain a noise analysis model.

[0034] Preferably, obtaining the denoised clean speech x1(n) of the first speech signal based on the first initial gain and the second initial gain includes the following steps:

[0035] The actual gain G(i,k) is obtained based on the first initial gain and the second initial gain.

[0036] G(i,k)=aG1(i,k)+bG2(i,k)+c

[0037]

[0038] In the formula, e is the natural constant, k is the better value after training, k∈(0,1], and a0 and b0 are both constants;

[0039] Among them, as the number of times the noise estimation is updated increases, the value of a decreases, the initial proportions are a = 0.5, b = 0.5, and c is the compensation coefficient;

[0040] The processed speech signal S(i,k) is obtained based on the actual gain G(i,k);

[0041] The speech signal S(i,k) is inversely transformed into the time-domain signal x1(n).

[0042] Preferably, the step of the speech noise reduction module in performing noise reduction processing on the second speech signal is the same as the step of the speech noise reduction module in performing noise reduction processing on the first speech signal.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention uses a voice acquisition module to acquire voice signals and sends the voice signals to a voice noise reduction module. The voice noise reduction module performs noise reduction processing on the first voice signal and the second voice signal respectively to obtain a first clean voice signal and a second clean voice signal. The first clean voice signal and the second clean voice signal are combined to obtain a clean voice signal. The noise reduction processing of the voice signal can reduce the amount of data communication in data communication and ensure the integrity of the data.

[0045] The AGC algorithm is used to compensate for clean speech signals. When the user's voice exceeds the acquisition limit, the gain is reversed to prevent audio distortion, thereby improving the fidelity and accuracy of the audio.

[0046] This invention places the voice acquisition module, voice noise reduction module, and voice recognition module on the mouse end for processing, ensuring the stability of the processing and eliminating the problem of distortion and poor noise reduction effect due to excessive data volume of the two audio data during transmission. It can use the original audio for noise reduction, thus improving the reliability of noise reduction. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the present invention;

[0048] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0049] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] like Figure 1-2As shown, a noise reduction system for a voice mouse includes a voice acquisition module, a voice noise reduction module, and a voice recognition module; the modules interact with each other based on digital signals; this noise reduction system is integrated into a voice processing chip located in the voice mouse body, and is used to process audio information at the mouse end to prevent audio distortion and loss during transmission when sending dual-channel audio.

[0051] The voice acquisition module is used to acquire voice signals; wherein, the voice signals include a first voice signal and a second voice signal;

[0052] And the voice signal is sent to the voice noise reduction module;

[0053] In this embodiment, the voice acquisition module includes a first voice acquisition unit and a second voice acquisition unit;

[0054] The first voice acquisition unit is used to acquire the first voice signal;

[0055] The second voice acquisition unit is used to acquire the second voice signal;

[0056] Specifically, both the first voice acquisition unit and the second voice acquisition unit acquire voice signals through a voice acquisition device;

[0057] The voice acquisition device includes a microphone;

[0058] It should be further noted that the microphone of the first voice acquisition unit is marked as the first microphone;

[0059] The microphone of the second voice acquisition unit is labeled as the second microphone;

[0060] The first microphone is installed on one side of the sound field, and the second microphone is installed on the same sound field but on the opposite side of the first microphone; the distance between the first microphone and the second microphone must be greater than 3 centimeters.

[0061] The speech noise reduction module is used to perform noise reduction processing on the speech signal;

[0062] Step 1: The speech noise reduction module performs noise reduction processing on the first speech signal and the second speech signal respectively to obtain a first clean speech signal x1(n) and a second clean speech signal x2(n);

[0063] In this embodiment, the speech noise reduction module performs noise reduction processing on the first speech signal, including the following steps:

[0064] S1: Extract the first M frames of the first speech signal as the first initial speech signal; where M is an integer greater than 0 and less than 20; it should be further explained that when collecting an audio signal, it is assumed that the first few frames of the noisy speech do not contain speech activity and are pure noise signals. Therefore, these first few frames can be used as the estimate of the noise spectrum.

[0065] S2: Input the first initial speech signal and the first speech signal data into the noise estimation model to obtain the first initial gain G1(i,k);

[0066] In this embodiment, the noise estimation model is used to obtain the gain G(i, k);

[0067] The principle of the noise estimation model includes:

[0068] y(n) = x(n) + d(n)

[0069] Where x(n) represents the clean speech signal, d(n) represents the noise signal, and y(n) represents the signal collected by the microphone;

[0070] Performing a short-time Fourier transform on y(n) = x(n) + d(n), we obtain:

[0071] Y(i,k)=X(i,k)+D(i,k)

[0072] Where Y(i,k), X(i,k) and D(i,k) represent the frequency domain signals of y(n), x(n) and d(n) respectively, i represents the i-th time domain frame, and k represents the frequency point;

[0073] G(i,k)=|X(i,k)| / |Y(i,k)|;

[0074] S3: Divide the first speech signal into P segments and count the energy of each segment;

[0075] The speech noise reduction module sets a quantile value and marks speech signals with energy below the quantile value as noise signals. It should be further explained that, based on a consensus, even in speech segments, the input signal may have no signal energy in certain frequency band components. The energy of all speech frames in a certain frequency band can be statistically analyzed, and a quantile value can be set. Energy below the quantile value is considered noise, and vice versa. This method, compared with frame-by-frame judgment, further refines the granularity of noise statistics, and can extract effective noise information even in speech frames.

[0076] S4: The speech noise reduction module inputs the speech signal marked as noise into the noise analysis model to obtain the noise probability; and marks the speech signal with the highest noise probability as the second initial speech signal;

[0077] In this embodiment, the noise analysis model is used to obtain noise probability; it should be further explained that noise classification features are obtained, and an artificial intelligence model is trained based on the noise classification features to obtain the noise analysis model;

[0078] S5: Input the second initial speech signal and the first speech signal data into the noise estimation model to obtain the second initial gain G2(i,k);

[0079] S6: Obtain the clean speech x1(n) after denoising the first speech signal based on the first initial gain and the second initial gain;

[0080] In this embodiment, obtaining the clean speech x1(n) after denoising the first speech signal based on the first initial gain and the second initial gain includes the following steps:

[0081] The actual gain G(i,k) is obtained based on the first initial gain and the second initial gain.

[0082] G(i,k)=aG1(i,k)+bG2(i,k)+c

[0083]

[0084] In the formula, e is the natural constant, k is the better value after training, k∈(0,1], and a0 and b0 are both constants;

[0085] Among them, as the number of times the noise estimation is updated increases, the value of a decreases, the initial proportions are a = 0.5, b = 0.5, and c is the compensation coefficient;

[0086] The processed speech signal S(i,k) is obtained based on the actual gain G(i,k);

[0087] The speech signal S(i,k) is converted into a time-domain signal x1(n) by inverse transform.

[0088] In this embodiment, the steps of the speech noise reduction module in processing the second speech signal are the same as the steps of the speech noise reduction module in processing the first speech signal.

[0089] After noise reduction, the system sends the processed audio data to the main control chip via a data bus inside the mouse, and then sends the data to the PC for recognition via 2.4G, USB or Bluetooth.

[0090] Step 2: Combine the first clean speech signal and the second clean speech signal to obtain the clean speech signal x(n);

[0091] Step 3: Compensate the clean speech signal according to the AGC algorithm;

[0092] In this embodiment, the clean speech signal is compensated according to the AGC algorithm, including the following steps:

[0093] The AGC algorithm is as follows:

[0094]

[0095] It should be further clarified that the 'a' in the formula is not related to the 'a' in obtaining the actual gain mentioned above;

[0096] Where A(n+1) is the gain value after signal transmission adjustment; A(n) is the gain value before adjustment; A in A0 is the signal amplitude value; A0 is the desired amplitude value, which is a constant.

[0097] Specifically, when the volume is below 58dB, the value of a is smaller and between (0, 5].

[0098] When the sound is [58, 75) dB, the value of a is between (12, 16);

[0099] When the sound is [75, 85] dB, the value of a is an infinite value to minimize gain;

[0100] When the sound level is greater than 85dB, gain 'a' needs to be changed to a negative value to reduce the audio amplitude.

[0101] The clean speech signal after compensation is x(n)A(n+1).

[0102] The speech recognition module is used to convert the noise-reduced speech signal into text data; it should be further noted that after the speech signal is converted into text data, it is used for semantic recognition and other operations.

[0103] Traditional audio noise reduction methods include:

[0104] The host computer collects audio from two microphones; the two audio streams are then combined.

[0105] The audio transmitted to the host computer is C1 = A1 + B1 (microphone 1).

[0106] C2 = A2 + B2 (Microphone 2)

[0107] C: Synthesized speech signal, A: User speech captured by microphone, B: Noise.

[0108] That is, the two audio streams transmitted to the host computer are user speech plus noise C1 collected by microphone 1, and user speech plus noise C2 collected by microphone 2.

[0109] The host computer will synthesize speech signals C1 and C2 to obtain a single speech signal C, identify a noise model signal B in a monostable model, and finally obtain the denoised speech A = CB for recognition.

[0110] The recognition results are improved compared to no noise reduction, but because the noise reduction requires compression when the audio signal is transmitted from the mouse to the host computer, some audio loss will occur.

[0111] This invention places the voice acquisition module, voice noise reduction module, and voice recognition module on the mouse end, ensuring the stability of the system's processing and eliminating the problem of distortion and poor noise reduction effect caused by excessive data volume in the two audio channels during transmission. It can use the original audio for noise reduction, improving its noise reduction reliability. The model has been optimized and modified, and an AGC algorithm model has been added to solve the feature loss caused by compression during the transmission of dual audio channels. Also, due to the addition of AGC, the volume of the audio signal can be adaptively adjusted, ensuring the recognition rate in both low and high volume situations.

[0112] In this embodiment, the voice acquisition module is communicatively and / or electrically connected to the voice noise reduction module;

[0113] The speech noise reduction module communicates with and / or is electrically connected to the speech recognition module.

[0114] The above formulas are all numerical calculations after removing dimensions. The formulas are obtained by software simulation based on a large amount of data and are closest to the real situation. The preset parameters and preset thresholds in the formulas are set by those skilled in the art according to the actual situation or obtained by simulation based on a large amount of data.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A noise reduction system for a voice mouse, characterized in that, It includes a voice acquisition module, a voice noise reduction module, and a voice recognition module; The voice acquisition module is used to acquire voice signals; wherein, the voice signals include a first voice signal and a second voice signal; And the voice signal is sent to the voice noise reduction module; The speech noise reduction module is used to process the speech signal to obtain a clean speech signal, including the following steps: Step 1: The speech noise reduction module performs noise reduction processing on the first speech signal and the second speech signal respectively to obtain the first clean speech signal x1(n) and the second clean speech signal x2(n); The speech noise reduction module performs noise reduction processing on the first speech signal, including the following steps: S1: Extract the first M frames of the first speech signal as the first initial speech signal; where M is an integer greater than 0 and less than 20; S2: Input the first initial speech signal and the data of the first speech signal into the noise estimation model to obtain the first initial gain G1(i,k0). S3: Divide the first speech signal into P segments and count the energy of each segment; The speech noise reduction module sets a quantile value and marks speech signals with energy below the quantile value as noise signals. S4: The speech noise reduction module inputs the speech signal marked as noise into the noise analysis model to obtain the noise probability; and marks the speech signal with the highest noise probability as the second initial speech signal; S5: Input the second initial speech signal and the first speech signal into the noise estimation model to obtain the second initial gain G2(i,k0); S6: Obtain the clean speech x1(n) after noise reduction of the first speech signal based on the first initial gain and the second initial gain; Obtaining the clean speech x1(n) after denoising the first speech signal based on the first initial gain and the second initial gain includes the following steps: The actual gain G(i,k0) is obtained based on the first initial gain and the second initial gain. G(i,k0)=aG1(i,k0)+bG2(i,k0)+c ,b=1-a; In the formula, e is the natural constant, k0 is the better value after training, k0∈(0,1], and a0 and b0 are both constants; Among them, as the number of times the noise estimation is updated increases, the value of a decreases, the initial proportions are a=0.5, b=0.5, and c is the compensation coefficient; The processed speech signal S(i,k0) is obtained based on the actual gain G(i,k0); The speech signal S(i,k0) is converted into a time-domain signal x1(n) by inverse transform. The steps of the speech noise reduction module in processing the second speech signal are the same as the steps of the speech noise reduction module in processing the first speech signal. Step 2: Combine the first clean speech signal and the second clean speech signal to obtain the clean speech signal x(n); Step 3: Compensate the clean speech signal according to the AGC algorithm; The speech recognition module is used to convert clean speech signals into text data.

2. The noise reduction system for a voice mouse according to claim 1, characterized in that, The voice acquisition module includes a first voice acquisition unit and a second voice acquisition unit; The first voice acquisition unit is used to acquire the first voice signal; The second voice acquisition unit is used to acquire the second voice signal.

3. A noise reduction system for a voice mouse according to claim 1, characterized in that, The noise estimation model is used to obtain the gain G(i,k); The principle of the noise estimation model includes: y(n) = x(n) + d(n) Where x(n) represents the clean speech signal, d(n) represents the noise signal, and y(n) represents the signal collected by the microphone; Performing a short-time Fourier transform on y(n) = x(n) + d(n), we obtain: Y(i,k)=X(i,k)+D(i,k) Where Y(i,k), X(i,k) and D(i,k) represent the frequency domain signals of y(n), x(n) and d(n) respectively, i represents the i-th time domain frame, and k represents the frequency point; G(i,k)=|X(i,k)| / |Y(i,k)|.

4. A noise reduction system for a voice mouse according to claim 1, characterized in that, An artificial intelligence model is trained based on noise classification characteristics to obtain a noise analysis model.

5. A noise reduction system for a voice mouse according to claim 1, characterized in that, The voice acquisition module, voice noise reduction module, and voice recognition module are located on the mouse.

Images